Build a generic ML algorithm to correct observed values
- or -
Post a project like this845
£363(approx. $455)
- Posted:
- Proposals: 17
- Remote
- #3465532
- OPPORTUNITY
- Awarded
191650651681116672338666634366048216255124561855352873585270488518946613009715109220
Description
Experience Level: Expert
We have conducted about 320 experiments. In each experiment, we measured X-values every 5 minutes and Y-values every 15 minutes. X is a single column that represents the observed values. Y is also a single column that represents the true values for any experiment.
The main objective of this project is to develop a generic machine-learning algorithm that can be used to correct X-values and make them as close as possible to Y-values for any experiment. The developed ML algorithm must account for the various deviations for X in these experiments.
Requirements:
To make the requirements as clear as possible let's assume that you have developed a generic model called (mdl) to correct X-values and Y_hat is the output of the model:
Y_hat =mdl(X-values);
1. The developed model will be applied in real-time applications using only X-values from any experiment so take into account that the model will output in real-time.
2. The final ML algorithm will be deployed as an android APK so take into account the complexity of the proposed model.
3. The mean absolute relative difference (MARD) metric must be used to assess the performance of the model. The target is to get MARD below 14% without using any true value in the model and below 12% if we use one true value from Y as an input to the model to calibrate the results and below 10% if we use two true values from Y as calibration points.
MARD = avreage(abs(Y_i - Y_hat_i)/Y_i)*100%
Notes:
1. Y (true values) are in this range [3 23], so the output of the model (Y_hat) should be in this range.
2. We can use the observed values in the first 2hours to derive features or detrend X-values and then update, tune and adjust the model parameters if needed where we can set the model to start generating outputs after 2hours.
3. It seems that X-values have a trend that needs to be removed. This trend sometimes is present only in the first 2hours so we need to take this into account. i.e. we can detrend the data to make it match the true values.
4. If the above can't be achieved then, we can also use only one true value or even two true values as input to the model to improve its performance by calibrating the final predicted true values if necessary.
Required Skills:
1. Data preparation and trends identification
2. Creative feature engineering
3. Advanced regression techniques
The main objective of this project is to develop a generic machine-learning algorithm that can be used to correct X-values and make them as close as possible to Y-values for any experiment. The developed ML algorithm must account for the various deviations for X in these experiments.
Requirements:
To make the requirements as clear as possible let's assume that you have developed a generic model called (mdl) to correct X-values and Y_hat is the output of the model:
Y_hat =mdl(X-values);
1. The developed model will be applied in real-time applications using only X-values from any experiment so take into account that the model will output in real-time.
2. The final ML algorithm will be deployed as an android APK so take into account the complexity of the proposed model.
3. The mean absolute relative difference (MARD) metric must be used to assess the performance of the model. The target is to get MARD below 14% without using any true value in the model and below 12% if we use one true value from Y as an input to the model to calibrate the results and below 10% if we use two true values from Y as calibration points.
MARD = avreage(abs(Y_i - Y_hat_i)/Y_i)*100%
Notes:
1. Y (true values) are in this range [3 23], so the output of the model (Y_hat) should be in this range.
2. We can use the observed values in the first 2hours to derive features or detrend X-values and then update, tune and adjust the model parameters if needed where we can set the model to start generating outputs after 2hours.
3. It seems that X-values have a trend that needs to be removed. This trend sometimes is present only in the first 2hours so we need to take this into account. i.e. we can detrend the data to make it match the true values.
4. If the above can't be achieved then, we can also use only one true value or even two true values as input to the model to improve its performance by calibrating the final predicted true values if necessary.
Required Skills:
1. Data preparation and trends identification
2. Creative feature engineering
3. Advanced regression techniques
Raed I.
0% (0)Projects Completed
-
Freelancers worked with
-
Projects awarded
17%
Last project
8 May 2024
United Kingdom
New Proposal
Login to your account and send a proposal now to get this project.
Log inClarification Board Ask a Question
-
There are no clarification messages.
We collect cookies to enable the proper functioning and security of our website, and to enhance your experience. By clicking on 'Accept All Cookies', you consent to the use of these cookies. You can change your 'Cookies Settings' at any time. For more information, please read ourCookie Policy
Cookie Settings
Accept All Cookies